distributed conceptual modelling in a swedish lowland...
TRANSCRIPT
-
318 © IWA Publishing 2013 Hydrology Research | 44.2 | 2013
Distributed conceptual modelling in a Swedish lowland
catchment: a multi-criteria model assessment
Sebastian Wrede, Jan Seibert and Stefan Uhlenbrook
ABSTRACT
Operational management and prediction of water quantity and quality often requires a spatially
meaningful simulation of environmental flows and storages at the catchment scale. In this study, the
performance of a fully distributed conceptual hydrologic model was evaluated based on the HBV
(Hydrologiska Byråns Vattenbalansavdelning) and TACD (Tracer Aided Catchment model – Distributed)
model concept in the meso-scale Fyrisån catchment in the Central Swedish lowlands. For a more
spatially explicit representation of runoff generation processes of small landscape elements such as
wetlands, a new sub-grid parameterization scheme was implemented in the model. In addition, a
simple flow distribution and lake retention routine was introduced to better conceptualize the flow
routing. During intensive model evaluation and comparison the model underwent conventional split-
sample and proxy-basin tests. In this process, shortcomings of the model in the transferability of
parameter sets and in the spatial representation of runoff generating processes were found. It was
also demonstrated how a detailed comparison with a lumped benchmark model and the additional
use of synoptic stream flow measurements allowed further insights into the model performance. It
could be concluded that such a thorough model assessment can help to detect shortcomings in the
spatial representation of the model and help facilitate model development.
doi: 10.2166/nh.2012.056
Sebastian Wrede (corresponding author)Stefan UhlenbrookFaculty of Civil Engineering and Geosciences,Water Resources Section,Delft University of Technology,NL-2600 GA Delft,The NetherlandsE-mail: [email protected]
Sebastian WredeCentre de Recherche Public – Gabriel Lippmann,Department of Environment and Agro-
Biotechnologies, L-4422 Belvaux,Luxembourg; andDepartment of Hydrology,University of Trier, D-54286 Trier,Germany
Jan SeibertDepartment of Geography,University of Zurich, CH-8057 Zurich,Switzerland;Department of Physical Geography and Quaternary
Geology, University of Stockholm,SE-10691 Stockholm,Sweden; andDepartment of Earth Sciences,Uppsala University, SE-75236 Uppsala,Sweden
Stefan UhlenbrookDepartment of Water Sciences and Engineering,UNESCO-IHE,NL-2601 DA Delft,The Netherlands
Key words | conceptual hydrological catchment models, distributed modelling, process-oriented
modelling, sub-grid variability
INTRODUCTION
Increasing environmental threats to water resources and
aquatic ecosystems call for further development of hydrologic
models that are able to better quantify environmental flows at
the catchment scale. However, this development has been
often associated with an increase in model complexity
along with the lack of observational data and appropriate
diagnostic tools to further constrain and evaluate model
states and outputs (e.g. Wagener et al. ; Gupta et al.
). Thus, appropriate model complexity needs to be
balanced with (i) the modelling purpose, (ii) the character-
istics of the hydrological system and (iii) the data available
(Wagener et al. ). In addition, powerful rigorous diagnos-
tic tests are needed to evaluate to what degree a realistic
representation of the natural system has been achieved and
how the given model concept can be improved (Gupta
et al. ). In this work a rigorous multi-criteria model
assessment was performed to evaluate a process-oriented, dis-
tributed hydrologic model that was developed to better
describe the spatial variability of hydrological states and
fluxes under boreal conditions in order to serve as a basis
for coupled solute transport applications (e.g. Lindgren
et al. ; Exbrayat et al. ).
Boreal landscapes are dominated by forest and open
land with distinct small scale landscape elements such as
lakes and wetlands that have a great influence on runoff
generation and solute transport (e.g. Arheimer & Wittgren
mailto:[email protected]
-
319 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
, ; Gren ). This mosaic of alternating land-
scape patches with individual characteristics needs to be
addressed by the chosen model concept. However, in most
cases computational limitations hinder fully resolving
spatial heterogeneity. Even distributed models are to some
degree spatially lumped and spatial elements are parameter-
ized using effective parameters. These parameters are
assumed to take into account spatial heterogeneity of land-
scape characteristics and hydrological processes within a
single model element, but might not be capable of reprodu-
cing the hydrological behaviour for an element as a sum
of its sub-elements. This is especially true when differences
in the functioning of the different sub-element units are sig-
nificant. Typical measures to account for this so-called ‘sub-
grid variability’ are commonly used in macro-scale appli-
cations (e.g. Blöschl & Sivapalan ). These can consist
of a statistical distribution function within a model element
or a process adequate areal discretization by subdividing
model elements into different sub-entities (Becker &
Braun ).
An additional element of distributed hydrologic model-
ling is the lateral routing of water along flow pathways
(surface and subsurface) and stream flow for which different
methods with varying complexity and data demand are
available (e.g. Singh ). Boreal environments are often
characterized by stream networks intersected by numerous
lakes that often lack detailed geometric descriptions. Conse-
quently, only very simple routing or flow distribution
functions are employed in conceptual modelling, such as
the triangular weighting function of the HBV (Hydrologiska
Byråns Vattenbalansavdelning) model (Bergström ).
Nevertheless, such simple approaches are not feasible in dis-
tributed model concepts, where more explicit weighting
functions accounting for spatial routing of flow through
the stream network are required.
Successful conceptual model applications depend on
accurate parameterization. This is usually achieved by com-
paring observed and modelled stream flow at the basin
outlet. This approach might be sufficient for simple
lumped models, but is not a rigorous enough criterion for
distributed model evaluation in order to ensure a correct
representation of internal state variables (e.g. Mroczkowski
et al. ). Additional information, such as groundwater
levels or soil moisture measurements, is required for a
sufficient multi-criteria calibration procedure, but avail-
ability of suitable data is generally poor or lacking in most
real-world applications. However, multi-scale validation by
including runoff series from different sub-catchments
enables an advanced parameter estimation and may lead
to a subsequent improvement of model consistency and per-
formance (Sooroshian & Gupta ).
To date, the most popular hydrological catchment
model in Scandinavia is the conceptual lumped (or semi-
distributed) HBV model (Bergström , ). It dates
back to the early 1970s and since then has been subject
to continuous improvements and developments resulting
in different model versions, such as HBV-96 (Lindström
et al. ), HBV-IWS (e.g. Hundecha & Bardossy ),
a Nordic HBV model version (Saelthun ) or a version
for nutrient modelling (HBV-N: Arheimer & Brandt
). For larger areas, such as the whole of Sweden, the
HBV model has been used in nested configurations.
More recently, raster based versions of the HBV model
(e.g. Beldring et al. ) were also developed. In such
applications model complexity needs to be balanced with
input data availability (Wagener et al. ), since
more complex conceptualizations naturally coincide with
increased parameter and model uncertainties (e.g. Beven
).
Here, a fully distributed, process-oriented catchment
model based on the HBV concept was evaluated for water
quality applications in a meso-scale lowland catchment in
Sweden with mixed land use. Research questions were: (i)
how to develop an efficient method to account for the sub-
grid variability of land use parameters within a raster-
based hydrological model; and (ii) how to integrate a
simple runoff routing routine for surface water bodies (chan-
nel network and lakes) in a lowland catchment? For
evaluation of the revised model concept an intensive
multi-criteria model assessment including synoptic runoff
data and a comparison with the standard lumped HBV
model was carried out.
THE FYRISÅN CATCHMENT
For model assessment in this study the meso-scale Fyrisån
catchment (Figure 1) was used. Several investigations have
-
Figure 1 | Fyrisån catchment and instrumentation network.
320 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
been carried out in this research area previously, including
the fundamental work of Hjulström () on the morpho-
logical activity of rivers. More recent studies dealt with
fluvial sediment transport and nutrient transport modelling
(e.g. Darracq et al. ; Lindgren et al. ) and the appli-
cation of various hydrologic catchment models (e.g. Seibert
, ; Motovilov et al. ; Xu ). The Fyrisån
catchment is situated in the eastern part of the central Swed-
ish lowlands, 60 km north of Stockholm. It belongs to the
Mälaren–Norrström drainage basin and covers an area of
approximately 2,000 km² before it discharges into Lake
Ekoln. The landscape is topographically and
morphologically characterized by the low-lying and flat Pre-
cambrian peneplain with most parts of the area ranging
between 30 and 50 m above sea level and a highest point
of 110 m. The area is mostly covered by forests (60%) of
pine and spruce or a smaller fraction of mixed deciduous
woodland, whereas the wide and flat river valleys in the
south, particularly around the city of Uppsala, are mainly
used for agriculture (32%). There are numerous wetlands
of varying extent which cover about 4% of the area. Lakes
constitute 2% of the landscape and are mainly small with
a mean surface area of 0.4 km² (Gretener ). Besides
the city of Uppsala, settlements (2%) are generally small in
-
321 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
extent and scattered over the area, and only a small portion
can be regarded as impervious. The distribution of predomi-
nant soil types can be roughly related to land use
information. Clay soils constitute most parts of the farmland
in the region, while till soils are generally covered by forest
(Seibert ).
The research area is located within a region of rela-
tively low annual precipitation with a corrected mean
annual precipitation for the meteorological station
Uppsala of 636 mm and an increase towards the west as
well as the east somewhat above 700 mm. Maximum pre-
cipitation is observed in August and minimum in February
and March. Snow constitutes about 20–30% to the total
precipitation with an average duration of snow cover of
100–110 days per year (Seibert ). The mean annual
temperature for the station Uppsala is þ6 WC with valuesranging from a maximum in June (þ17 WC) to a minimumin February (–5 WC). The runoff regime shows a typical
Baltic regime with a dominant primary snowmelt spring
flood in April, a secondary rainfall peak flow in autumn
and a low flow period during the summer months.
MODEL DEVELOPMENT
Lumped HBV model
The conceptual lumped runoff model used in this study was
the HBV model (Bergström , ). Daily discharge is
simulated by using daily precipitation and temperature
data as well as monthly estimates of potential evaporation
as driving variables. The conceptualization of hydrological
processes employs sequentially linked routines and func-
tions representing the major processes of the land phase
hydrological cycle. These routines include a snow module,
simulating snowmelt with the degree-day method, a soil rou-
tine where groundwater recharge and actual evaporation are
functions of actual water storage in a soil box, and a runoff
generation routine where runoff from the groundwater stor-
age is represented by linear storage equations. Channel
routing is simulated via a triangular weighting function.
Detailed descriptions of the basic HBV model including
the governing equations can be found elsewhere (e.g.
Bergström ; Seibert ).
Distributed HBV model
In this work the lumped HBV model has been extended to a
fully distributed (grid based) version with a spatial resolu-
tion of 250 × 250 m2. While the general model structure
from the lumped HBV model remained unchanged, key
elements for the spatial distribution of the model were
adopted from the distributed and process-oriented catch-
ment model TACD (Tracer Aided Catchment model –
Distributed) (Uhlenbrook et al. ; Uhlenbrook &
Sieber ). These elements contain the modular model
structure with adapted runoff generation routines as well
as the integration into the geographical information system
PCRaster (Karssenberg et al. ). PCRaster offers a
dynamic modelling language for raster based applications
and enables lateral cell to cell routing with a single-flow
direction algorithm (D8) (O’Callaghan & Mark ). This
distributed and revised version of the HBV model, including
its modifications, aims at providing a more process-oriented
and spatially more explicit representation of the hydrologi-
cal system under boreal conditions. It further serves as a
hydrological basis for the solute transport model HBV-ND
that has been used in several studies (Lindgren et al. ;
Exbrayat et al. ).
In the following paragraphs two major new model
elements, which were added to the previous model versions
of HBV or TACD and were tested in the lowland Fyrisån
catchment, are described in more detail. These are an
approach to consider sub-grid variability and a distributed
flow routing method.
Sub-grid variability
The new sub-grid variability scheme using fractions of land-
use classes rather than the usual approach to assign the
major class to the entire grid cell has several advantages.
In particular, it is effective when the model grid cell resolu-
tion exceeds the resolution of comparatively smaller scale
land use patterns or landscape features, as often found in
boreal landscapes, so that their accurate representation is
no longer feasible. A conventional procedure to counteract
this problem consists of an increased spatial resolution by
decreasing the grid cell size. However, this is usually not
a practical solution as a better representation of small
-
Figure 2 | Schematic representation of the sub-grid parameterization scheme in a singlemodel grid cell. All water flows and storages within the model routines are
weighted according to their designated land use class. Vertical flows indicate
fluxes between different model storage compartments. Lateral flows indicate
fluxes to neighbouring draining grid cells.
Table 1 | Land use distribution depending on grid resolution for the Fyrisån catchment.For the conventional aggregation the major land use class was assigned to
the entire cell
Original data set (25 × 25 m2)and sub-grid parameterization(250 × 250 m2)
Conventionalaggregation(250 × 250 m2)
Land use km² % km²% oforiginal
Catchmentsize
2,006 100 2,063 102.9
Forest 1,202 59.9 1,303 108.4
Agriculture 646 32.2 638 98.8
Wetlands 91 4.5 57 62.6
Settlements 34 1.7 35 102.9
Lakes 32 1.6 31 96.9
322 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
scale landscape features with such a high spatial resolution
results in a considerable rise of model computation time.
Strasser & Etchevers () provided an example that
necessitated an increase in grid cell resolution of elevation
and meteorological input data by a factor of 64 to improve
variability of snowmelt in model simulations. They argued
that sub-grid parameterization methods are beneficial in
order to achieve reasonable computation time and to
fulfil data requirements. As such, a sub-grid parameteriza-
tion scheme certainly has clear advantages. Computation
time is comparably smaller than for higher spatial resolu-
tions, as it is dependent only on the number of different
land use classes or landscape features. Moreover, sub-grid
parameterization is area accurate, meaning that the grid
resolution has no effect on the correct representation of
surface area fractions.
Similar to the approach used in TACD, runoff generation
was parameterized separately for distinct landscape
elements, such as land use types. However, this traditional
approach of attributing the major land use type in a grid
cell to the entire cell becomes problematic when larger
grid cell sizes are used, as smaller landscape elements
might not be correctly represented or even neglected. There-
fore, a new sub-grid parameterization scheme was
introduced, where different land use types within a single
grid cell are represented by their fraction of the entire cell.
In the present model, five land use classes (forest, agricul-
ture, urban area, lakes and wetlands) are distinguished,
each representing a specific runoff generation routine. The
relative areas of these five classes are then used to weight
flow and storage amounts and to compute mean flow and
storage conditions for the entire grid cell (Figure 2). Lateral
cell to cell processes include flow from the groundwater sto-
rage to neighbouring cells as well as channel routing.
In the case of the Fyrisån catchment, conventional grid
aggregation led to a substantial decrease of small scale
landscape features in the model representation, as illus-
trated by the wetland area which was underestimated by
almost 50% (Table 1). This dramatic decline of the esti-
mated wetland area can be explained by the patchy and
small scale character of wetlands in this area that cannot
be captured by the coarse raster structure and even results
in an overestimation of catchment and forest area during
grid aggregation.
Simulation of channel routing and lakes
In distributed modelling, flow routing through channel net-
works and lakes plays an essential role in larger scale
-
323 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
model applications. In the Fyrisån catchment conventional
routing methods, such as the kinematic wave approach
applied in the TACD model (Uhlenbrook et al. ), were
not applicable as detailed information on channel and
lake properties was lacking. Also, the application of much
simpler techniques such as the MAXBAS weighting function
of the HBV model was neglected in favour of a spatially
more explicit representation. A simple method accounting
both for spatial and temporal distribution of water flow
within the channel network was needed. A relatively
simple routing module was developed that computes a
downstream distribution of water content for each stream
grid cell, according to a parameterized triangular weighting
function. Water fractions from a stream grid cell are distrib-
uted over its adjacent downstream cells (Figure 3). The form
of the weighting function is dependent on two parameters
DMAX [-] and DPEAK [-]. DMAX specifies the number of
downstream grid cells over which the water content of the
initial grid cell spreads, while DPEAK is a measure for the
location of the maximum peak of the triangular weighting
function, as shown in Figure 3. This conceptualization
allows a more flexible parameterization of the flow routing
in complex channel networks by allowing symmetric (com-
parable to MAXBAS in the traditional HBV model) and
non-symmetric flow weighting and distribution. A daily
time step in the implementation of the spatial distribution func-
tion appeared adequate to capture the temporal propagation of
water flow. However, this channel flow is interrupted by
numerous lakes in which flow retention occurs. Thus, these
Figure 3 | Downstream distribution of stream water according to the flow distributionparameters DMAX and DPEAK of the flow routing routine.
lakes are explicitly accounted for as separate storage reservoirs
in the model structure and the daily stream discharge at lake
inlets is attributed to these lake storages in conjunction with
lateral inflows of adjacent land use cells. At the respective
lake outlets, outflow is computed on the basis of a nonlinear
power function depending on inflow and water storage. After-
wards, the outflow is added to the next stream section, where
the downstream distribution of water continues. This
implementation of flow routing in combination with the
spatially explicit representation of lakes throughout the catch-
ment allows a more process realistic conceptualization of lake
retention and storage with only two parameters.
DATA BASE AND MODELLING PROCEDURE
Spatial data
The low topographic gradient throughout the lowland Fyri-
sån watershed was one main concern during data
preparation. For the correct delineation of sub-catchments
and the local drainage network a high resolution digital
elevation model (DEM) would have been beneficial, but
was unfortunately not available. Therefore, digital elevation
data, derived by the Shuttle Radar Topography Mission
(SRTM) at a 3 arc-seconds 90 × 90 m resolution (SRTM-3)
was used (e.g. Rabus et al. ) and resampled to the
250 × 250 m2 resolution of the model domain.
The local drainage network delineation according to the
D8 routing algorithm (O’Callaghan &Mark ) was based
on the SRTM DEM data set. Due to the low topographical
relief in this part of Sweden it was necessary to consider
auxiliary information of a predefined stream network to
force the computed drainage network through this observed
stream network (Hutchinson ).
Information about land use, including stream network
and lakes, was available as a vector data set from respective
topographical maps (Lantmäteriet ). Unfortunately, no
detailed soil map was available for the research area. How-
ever, the major land use classes can be seen as a surrogate
for certain dominating soil classes; in forested areas till
soils are common whereas agricultural areas correspond
to clay soils (Seibert ). Five characteristic hydrological
response units (HRUs) were differentiated (forest,
-
324 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
agriculture, wetland, settlement, and lake) based on land use
information. These five different HRUs were derived by re-
classification and aggregation of 23 available land use
classes and parameterized individually with the help of typi-
cal literature values about the prevailing soil characteristics,
to allow for the better characterization of different major
runoff generation processes in each HRU.
Meteorological and hydrological data
Meteorological and hydrological time series for the investi-
gation period from 1992 to 2005 were obtained from the
standard observation network of the Swedish Meteorological
and Hydrological Institute (SMHI). Daily uncorrected pre-
cipitation data at eight stations situated within the vicinity of
the research area and northwards from Uppsala were avail-
able. Moreover, three climate stations provided daily mean
temperature data. Inverse distance weighting was used for
regionalization of meteorological input data. Eriksson ()
published monthly potential evaporation estimates for the
whole of Sweden from which suitable values for the River
Fyris were selected. The hydrological discharge observation
network from SMHI contains three regular stations within
the drainage basin (Figure 1), but no outlet station capturing
the total drainage basin was available. The sub-catchments
Vattholma (281 km²) and Sävja (717 km²) had continuous
records over thewhole application periodwhile for the station
of Ulva Kvarn (950 km²) (including the smaller Vattholma
sub-catchment) discharge measurements exist only until the
year 2000. To support the model application additional data
was collected by a synoptic dischargemeasurement campaign
that was carried out during low flow at 26 locations within the
Fyrisån drainage basin on 29 June 2005.
Model parameterization and calibration
In this study an automated model calibration procedure was
applied by coupling the hydrologic model to the parameter
estimation program PEST (Parameter ESTimation). PEST
is a model-independent nonlinear parameter estimation
and optimization package, frequently used for model calib-
ration in different research fields (Doherty & Johnston
; Doherty ). It is based on the implementation of
the Gauss–Marquardt–Levenberg algorithm, which
combines the advantages of the inverse Hessian method
and the steepest gradient method to allow a fast and efficient
convergence towards the objective function minimum. The
best model parameter set is selected within a specified
range of parameter values by minimizing the discrepancies
between model results and simulated or predefined values
in a weighted least square sense.
For the automatic model calibration, 19 out of 28 model
parameters were selected. The remaining parameters were
fixed according to literature values or tied with an equal
ratio to preceding parameters that were subject of the calib-
ration process in order to reduce the overall parameter
space and enable a fast and successful calibration result
(Table 2). The relatively large amount of parameters results
from the separate parameterization of runoff generation for
each different land use class. A detailed sensitivity and uncer-
tainty analysis could not be achieved in the framework of this
paper as it was computationally too demanding. However, in
a similar model concept with even higher parameterization,
Sieber & Uhlenbrook () conducted a sensitivity analysis
of model parameters by using the GLUE (Generalized Likeli-
hood Uncertainty Estimation) approach and demonstrated
the plausibility of themodel structure and process conceptual-
izations. This study, in combination with earlier HBV model
applications (e.g. Bergström ; Seibert , ; Uhlen-
brook et al. ) as well as TACD model applications (e.g.
Ott & Uhlenbrook ; Uhlenbrook et al. ; Wissmeier
& Uhlenbrook ; Johst et al. ), guided the selection
of upper and lower bounds for each parameter of the distrib-
uted HBV model. Initial parameter values were selected on
the basis of previous best manual calibration trials in order
to start with the optimal known parameter set. By choosing
these initial parameter sets from different ranges within the
parameter space, the capability of PEST to find the global
minimum over local minima of the objective function by the
calibration algorithm was determined.
To evaluate the performance of the obtained parameter
sets in the course of the automatic calibration process, differ-
ent objective functions were used as they judge the model
performance by different aspects (Seibert ). In addition
to model efficiency Reff (Nash & Sutcliffe ) and volume
error VE criterions, the RV criterion proposed by Lindström
et al. () were computed (Table 3). The latter was finally
chosen for automatic model calibration, as it adequately
-
Table 2 | Model parameters with ranges and initial values used for the PEST calibration of the hydrological model
Parameter Explanation Unit Initial Minimum Maximum Estimate
Snow routine
TT Threshold temperature WC 0 –2.5 2.5 Calibrated
TTdiff TT for forestWC 0 –2.5 2.5 Calibrated
SFCF Snowfall correction factor – 0.6 0.4 1 Calibrated
SFCFdiff SFCF for forest – 0 0.4 1 Calibrated
CFMAX Degree-day factor mm WC–1 d–1 2 1 8 Calibrated
CWH Water holding capacity – 0.1 – – Fixeda
CFR Refreezing coefficient – 0.05 – – Fixeda
Soil routine
LP Reduction of evaporation – 0.6 0.3 1 Calibrated
FCforest Field capacity for forest mm 300 50 500 Calibrated
FCagricul Field capacity for agriculture mm 200 50 500 Calibrated
FCwetland Field capacity for wetland mm 100 – – Tied
FCurban Field capacity for urban mm 150 – – Tied
BETAforest Shape coefficient for forest – 4 1 6 Calibrated
BETAagricul Shape coefficient for agriculture – 4 – – Tied
BETAwetland Shape coefficient for wetland – 4 – – Tied
BETAurban Shape coefficient for urban – 4 – – Tied
Runoff generation routine
UrbanSplit Portion of sealed urban areas d–1 0.5 – – Fixed
KUS forest Upper recession coefficient for forest d–1 0.25 0.01 0.4 Calibrated
KLS forest Lower recession coefficient for forest d–1 0.005 0.001 0.15 Calibrated
PERCforest Percolation from upper to lower box forest mm d–1 0.05 0.001 3 Calibrated
HUS forest Maximal storage capacity upper box forest mm 350 1 1,000 Calibrated
HLS forest Minimal storage capacity lower box forest mm 1,000 – – Fixed
KUS agricul Upper recession coefficient for agriculture d–1 0.35 0.01 0.4 Calibrated
KLS agricul Lower recession coefficient for agriculture d–1 0.007 0.001 0.15 Calibrated
PERCagricul Percolation from upper to lower box agriculture mm d–1 0.008 0.001 3 Calibrated
HUS agricul Maximal storage capacity upper box agriculture mm 250 1 1,000 Calibrated
HLS agricul Minimal storage capacity lower box agriculture mm 1,000 – – Fixed
KUS wetland Upper recession coefficient for wetland d–1 0.05 0.01 0.4 Calibrated
HUS wetland Maximal storage capacity upper box wetland mm 150 – – Calibrated
KUS urban Upper recession coefficient for urban d–1 0.5 0.01 0.4 Calibrated
KLS urban Lower recession coefficient for urban d–1 0.003 0.001 0.15 Calibrated
PERCurban Percolation from upper to lower box urban mm d–1 0.01 0.001 3 Calibrated
HUS urban Maximal storage capacity upper box urban mm 100 1 1,000 Calibrated
HLS urban Minimal storage capacity lower box urban mm 1,000 – – Fixed
Lake and flow distribution routine
Klake Recession coefficient lake d–1 1 0.001 1 Calibrated
ALPHAlake Nonlinear weighting coefficient – 1 0.001 1 Calibrated
DMAX Flow distribution length – 109.1 3 160 Calibrated
PEAK Flow distribution peak location – 81.83 – – Tied
aBergström (1995).
325 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
-
Table 3 | Objective functions
Objectivefunction Symbol Definition Unit
Value for‘perfect’fit
Efficiencya Reff 1�Pn
i¼1 Qi,obs �Qi,sim� �2
Pni¼1 Qi,obs �Qobs
� �2 – 1
Relativevolumeerror
VE
Pni¼1 Qi,obs �Qi,sim
� �
Pni¼1 Qi,obs
– 0
RV criterionb Rv Reff �wjVEj – 1
aNash & Sutcliffe (1970).bLindström et al. (1997) with weight: w¼ 0.1.
326 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
balanced model efficiency Reff and relative volume error VEof the model.
Model calibration against daily runoff was carried out for
two sub-catchments, Vattholma and Sävja, with the inclusion
of the runoff station Ulva Kvarn for validation purposes. The
simulation period of this study ranged from October 1994 to
June 2005, preceded by a 2 yearwarming-up period starting in
October 1992. This period was divided into two sub-periods
of 5 years, starting with the calibration period from October
1994 to September 1999 and followed by a validation
period from October 1999 to June 2005.
Model validation was conducted step-wise according to
the hierarchical scheme for systematic testing of hydrological
simulation models proposed by Klemes (): In a first step,
the model was calibrated for both sub-catchments individu-
ally. Afterwards, the best derived parameter set for each
catchment was exchanged and used for a new model simu-
lation in the adjacent catchment, employing the proxy-
basin test procedure. In a third step, themodel was calibrated
on both catchments in order to determine an optimal joined
parameter set and, in a last step, all derived parameter sets
underwent a traditional split-sample test procedure.
RESULTS
Split-sample test results
The results of the individual catchment calibration showed
satisfactory fits between measured and simulated discharge
for the calibration period with RV values ranging from
0.85 to 0.90. Noticeable is a strong decline during the vali-
dation period from 0.90 to 0.73 (RV) for Vattholma
compared to 0.85 to 0.77 (RV) for Sävja (Table 4).
A similar model performance was observed for the vali-
dation period using the individual optimized parameter sets
in each of the other watersheds. While RV values for the first
5 years remained at a reasonable level of 0.73 and 0.76, the
following 5 years showed significantly poorer statistical
measures with the strongest declines for the Vattholma
sub-catchment. In addition to this decrease in efficiency
(RV) for the second half of the time period the relative
volume error increased and revealed, along with RV, a sys-
tematic underestimation of the flow dynamics and volume
by the model. These apparent errors of the model simulation
became evident for single years 1996 and 1997 of the calib-
ration period, but also dominated throughout the second
half of the Vattholma runoff record, especially in 2000 and
2004, where years with multiple spring melt peaks prevailed.
Figure 4 demonstrates the lack of the model to capture the
variable runoff situation for this period with almost opposite
simulations of the flow dynamics. In contrast to this, the
model performances for the remaining years are satisfactory
and are shown for the hydrological year 1998 in Figure 5. In
this case the model is able to capture quite accurately the
entire runoff dynamics on the basis of the individual
Vattholma parameter set.
Proxy-basin test results
It is interesting to further evaluate the hydrologic differences
between the two catchments. Therefore, efficiency values
(Reff) were computed as a benchmark for comparing (i)
the observed specific runoff records from Sävja to Vat-
tholma and (ii) the simulated specific runoff records from
these basins as it was proposed by Seibert et al. ().
Results are listed in Table 5 and reveal greater similarities
(higher Reff) between the model generated time series than
between the measured specific discharges (lower Reff). It is
also apparent that the model performed better (Table 4)
than the benchmark of simply transferring the specific dis-
charge between the two catchments (Table 5).
The aspect of model parameter dependency on individ-
ual catchments was further tested with simultaneous
calibration on both catchments to derive a joined parameter
-
Figure 4 | Hydrological year 2000, which was the year with the worst model fit forVattholma.
Figure 5 | Hydrological year 1998, which was the year with the best model fit forVattholma.
Table 4 | Overview of statistical performance measures for split-sample and proxy-basin test results. Simulation performance for data used in the calibration period is shown in bold, forthe validation period in regular font: calibration period (01.10.94–30.09.99); validation period (01.10.99–30.06.05)
Vattholma Sävja Ulva Kvarn
Catchment used for calibration Objective function 1994–1999 1999–2005 1994–1999 1999–2005 1994–1999
Vattholma (281 km2) RV 0.90 0.73 0.73 0.63 0.79Reff 0.90 0.73 0.75 0.65 0.81VE 0.00 0.03 0.17 0.16 0.18
Sävja (717 km2) RV 0.76 0.54 0.85 0.77 0.83Reff 0.78 0.57 0.85 0.78 0.83VE 0.19 0.25 0.02 0.06 0.01
Vattholma & Sävja ( joined parameter set) RV 0.83 0.63 0.83 0.77 0.82Reff 0.84 0.64 0.84 0.78 0.83VE 0.09 0.13 0.10 0.07 0.11
327 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
set. This was achieved by assigning the same weights (50%)
to the objective functions for each catchment. It led to a gen-
eral decrease of model performance for the calibration
phase compared to the results obtained by calibration on
individual catchments alone, but provided an increase of
the overall model performance for the whole application
period. Validation also included the runoff station Ulva
Kvarn with independent runoff records of the calibration
phase. When comparing the statistical performance
measures for all these catchments in Table 4, it becomes evi-
dent that the overall best fit could be achieved with the
joined parameter set. This was the best parameter set that
could be derived in this study for a model application cover-
ing the whole Fyrisån basin.
-
Table 5 | Comparison of efficiencies (Reff) derived from runoff records from Vattholmaand Sävja
Calculation of efficiency between Vattholmaand Sävja based on…
Resulting efficiency(Reff)a
Observed specific runoff records 0.63
Simulated specific runoff records 0.84
aNash & Sutcliffe (1970).
328 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
Model comparison
The performance of the distributed model was compared to
the lumped HBV model using the same regionalized input
data (precipitation and temperature as well as monthly
potential evaporation estimates) for both model appli-
cations. In this case it becomes evident that the
distributed, highly parameterized model was not able to out-
perform the simpler, less parameterized lumped HBVmodel
in terms of runoff related efficiency measures (Table 6). For
most years similar model behaviour can be observed result-
ing in model errors for the same years throughout all
catchments (Figures 4 and 5).
Synoptic runoff measurements
Synoptic runoff measurements from a field campaign in the
River Fyris allowed the distributed model simulations to be
evaluated at several grid cells along the stream network
during low flow conditions. These synoptic runoff measure-
ments were used in two ways. In a first step observed and
simulated runoff volumes were compared. Figure 6 illus-
trates that, in general, flow volumes could be reproduced
reasonably well by the distributed model. The second evalu-
ation compared observations and simulations as specific
runoff values, i.e. runoff divided by the respective sub-
catchment areas. This latter evaluation removed the
Table 6 | Comparison of distributed vs. lumped model results obtained with joined parameter
Objective function Vattholma S1994–1999 1999–2005 1
RV 0.83 0.86 0.63 0.74 0
Reff 0.84 0.86 0.64 0.76 0
VE 0.09 0.04 0.13 0.01 0
influence of catchment size and resulted in large scatter in
Figure 7, indicating shortcomings of the distributed model
in the simulations of spatial variations. The additional
inclusion of these synoptic measurements in the calibration
only partly improved the performance (Figure 7).
DISCUSSION
Split-sample test results
In general, the model performance was acceptable for calib-
ration and validation periods for the catchments for which
the model had been calibrated. Model efficiencies were
smaller for the validation periods, which can partly be
explained by a clear change in runoff regime from one
dominated by a single spring flow to a more erratic unstable
flow regime with multiple snow-melt runoff events. Unstable
flow regimes have been noted during former model appli-
cations in this region (Motovilov et al. ) and have
been further investigated by Krasovskaia & Gottschalk
() for Scandinavian countries.
Proxy-basin test results
The proxy-basin test revealed additional shortcomings in the
spatial representation of the model. Despite the inclusion of
the spatially explicit, land use dependent runoff generation
routine and the distributed flow and lake routing, the model
was not able to capture major changes in runoff for different
catchments on the basis of the individually calibrated para-
meter set (Table 4). This was reflected in low efficiencies
(RV) for the exchanged parameter sets, especially for the
second part of the application period. Once again differences
in runoff regimemight come into play, but computed efficien-
cies (RV) of the measured as well as simulated specific
set. Results from distributed model (bold); results from lumped model (regular)
ävja Ulva Kvarn994–1999 1999–2005 1994–1999
.84 0.86 0.78 0.79 0.82 0.84
.84 0.86 0.78 0.79 0.83 0.84
.1 0.04 0.07 0.01 0.11 0.05
-
Figure 7 | Validation of the model performance by comparison of synoptic specificdischarge measurements with simulated synoptic specific discharge
(uncalibrated: r¼ 0.22; calibrated on synoptic measurements: r¼ 0.36).
Figure 6 | Validation of the model performance by comparison of synoptic dischargemeasurements with simulated discharge (r¼ 0.97).
329 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
discharges of each catchment indicate a more basic problem:
while efficiencies (RV) for the transferred parameter sets are
low, signifying considerably different runoff behaviour of the
two sub-catchments, the calibrated parameter sets provide
much higher efficiencies (RV) and thereby show that the
diverse runoff character was not sufficiently met by the
model structure. This is in line with a similar proxy-basin
test study of TOPMODEL, where Donnelly-Makowecki &
Moore () also found to their surprise that their model
that explicitly accounted for topography was not superior
to a lumpedmodel when transposing it to another catchment.
The fact that parameter sets adapted to one catchment
cannot be simply transferred to the adjacent catchments,
although the most important spatial processes controlling
the flow regime are included in the model structure, under-
lines the effective character of these parameter sets that still
partly incorporate regional spatial heterogeneity character-
istics of each catchment. Nevertheless, it should be also
stated that proxy-basin tests often result in rather drastic per-
formance reductions and are even failed by many established
models (Refsgaard&Henriksen ). This might also be the
reason why it has been rarely applied, despite its rather infor-
mative character (Andreassian et al. ).
The simultaneous calibration of the model to both catch-
ments supported this finding with the presence of a joined
parameter set that is able to adequately capture the entire
runoff hydrographs for all sub-catchments. This joined para-
meter set accounts for regional runoff dynamics with a
slightly reduced efficiency (RV), but performs much better
on an overall basis than both previous individual parameter
sets (Table 4).
Besides the model application another often applied
approach for runoff predictions in ungauged basins is the
transfer of specific discharge from a nearby watershed
scaled by the catchment size. This alternative reveals
mostly convincing results for catchments with almost identi-
cal input data (e.g. Seibert et al. ) and was compared to
the prior model outputs. Table 4 reveals significantly higher
efficiencies (Reff) of transferred model parameters against a
considerably reduced Reff of the simple transfer method
(Table 5). As both basins are located in close vicinity, both
are subject to almost identical meteorological conditions,
so that the differences in land use and lake distribution are
assumed to be the reason for the better model performance.
This reflects the value of model applications compared to
simpler alternatives, despite the aforementioned deficits to
adequately account for spatial heterogeneity.
Besides the model structure, another point in the discus-
sion that should be considered is the applied calibration
method. Simple lumped models do not suffer from high com-
putation times, so intensive calibration procedures which
necessitate extensive model runs (e.g. Monte Carlo simu-
lations or genetic algorithms) can be easily employed. On
the other hand, distributed models mostly lack efficiency in
computation time and exhibit, in most cases, even higher
-
330 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
parameterizations due to complex spatial structures. There-
fore, less model runs for calibration purposes are feasible
and the optimal global parameter set is not inevitably achieved
every time. This effect might be underestimated, but could be
clearly verified on the basis of different start parameter sets for
the coupled parameter estimator PEST. It was found that the
variation of initial parameters resulted in different optimized
parameter sets with varying model performances.
Model comparison
By further comparing the lumped HBV model with the dis-
tributed model, it was shown that the latter did not lead to a
significant improvement of the discharge simulation per-
formance. Both models were more or less equivalent in
their success in simulating discharge at the different catch-
ment outlets and performed equally well in split-sample
and proxy-basin tests.
However, such equal model performance was not necess-
arily anticipated beforehand. Concerning the transfer of
model parameters to adjacent catchments, it was expected
that the distributed and more process-oriented model would
outperform the lumped concept, based on its spatially more
explicit representation of hydrological processes and its
increased degrees of freedom resulting from the enlarged
parameter space. Within the old debate about the value of
distributed versus lumped modelling (e.g. Beven ;
Refsgaard et al. ) this model comparison may thus be
seen as another example for the supremacy of less parameter-
ized lumped model concepts, if the objective is the best
discharge fit at the catchments outlet, given a limited input
data set that is explored to its full potential (e.g. Jakeman &
Hornberger ; Lischeid & Uhlenbrook ; Carpenter &
Georgakakos ; Breuer et al. ). This confirms, in
accordance with earlier work (e.g. Grayson et al. ;
Michaud & Sorooshian ; Refsgaard & Knudsen ;
Beven ), that distributed models seldom demonstrate
superiority over much simpler lumped or semi-distributed
models, if tested only against runoff at the catchment outlet
that constitutes lumped data, integrated over the whole catch-
ment. In contrast, themain advantage of a distributedmodel is
its capability of simulating additional internal state variables
that can be subject to multi-criteria calibration, if additional
data is available (e.g. groundwater data, synoptic runoff
measurements). This is especially important in terms of evalu-
ating model structure uncertainty (Fenicia et al. ) and
refining hydrological process descriptions in order to prevent
the ‘the model is right for the wrong reasons’ case (Klemes
), to which lumped model concepts may tend to.
Such spatially differentiated and more process-oriented
simulations might be beneficial in integrated catchment
modelling when the focus lies on the more accurate rep-
resentation of spatial patterns. For the case of nutrient
transport modelling, point sources and diffuse sources
could be incorporated according to their spatial represen-
tation so that degradation and retention patterns can be
simulated in a spatially more representative way (e.g.
Lindgren et al. ). In situations where models with
rather coarse grid resolutions need to consider small scale
processes in an adequate manner with limited computation
power, a sub-grid parameterization can be appropriate. This
is frequently the case at larger scales such as regional cli-
mate models with land surface schemes (e.g. Kotlarski &
Jacob ) or Soil-Vegetation-Atmosphere-Transfer (SVAT)
models (e.g. Strasser & Etchevers ), but can also be
transferred to meso-scale model applications like nutrient
transport modelling (e.g. Exbrayat et al. ), where small
lakes, wetlands or riparian zones can have a considerable
impact on nutrient flows and distributions (Gren ;
Carpenter et al. ; Hooper ).
Synoptic runoff measurements
The synoptic runoff measurements allowed an evaluation of
the spatial representativity of the model simulation for dis-
tributed runoff predictions. Overall, the model was able to
reproduce the spatial variations of runoff volumes, but not
the variations in specific runoff. This multi-scale evaluation
indicated that the model can capture scale induced runoff
volumes, but was not capable of reproducing the spatial vari-
ations of runoff generation processes in a low flow situation,
despite all calibration efforts. This might be partially
explained by measurement uncertainties and shortcomings
to correctly delineate sub-catchments in this lowland
region. In addition, potential evaporation estimates were
not differentiated for different land use classes which
might further explain these shortcomings. However, using
only one snapshot campaign during low flow conditions is
-
331 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
certainly the most important limitation, but the poor agree-
ment between simulated and predicted specific runoff gave
an indication that a further evaluation of the spatial rep-
resentation of runoff generating mechanisms in the model
structure is needed. This corresponds to other authors that
argue for spatially distributed data to fully evaluate the capa-
bilities of a distributed model (Gupta et al. ; Grayson &
Blöschl ). Despite such limitations, synoptic runoff
measurements can provide useful insights for model evalu-
ation or calibration purposes with a few additional
observations (e.g. Perrin et al. ; Seibert & Beven ).
CONCLUSION
In this paper a fully distributed, conceptual hydrological
model was evaluated that can serve as a basis for water qual-
ity applications in the Fyrisån catchment in Sweden. To
enable a better representation of small scale landscape fea-
tures and related runoff generation processes, such as
wetlands, a sub-grid parameterization scheme was incorpor-
ated into the model. Flow routing along stream networks
and lakes was achieved by a new designed simple flow rout-
ing routine. The objective of the presented model was to find
a spatially meaningful hydrological model that can provide
driving variables for coupled distributed solute transport
routines (Lindgren et al. ; Exbrayat et al. ).
Intensive model assessment and comparison throughout
the study revealed limitations of the model capabilities,
especially with respect to the transferability of model para-
meters and the simulation of spatial runoff patterns. The
model performed only equally well compared to the much
simpler lumped HBV model with regard to simulating
runoff at the catchment outlets. However, it is important
to note that the identification of these model shortcomings
was only possible due to the rigorous tests that were carried
out during the model evaluation process and once again
highlight the importance of thorough model evaluation pro-
cedures (Andreassian et al. ). Besides the well-known,
albeit less widely applied, test procedures suggested by
Klemes (), it was demonstrated that the comparison
with hydrologic benchmark models, such as the lumped
HBV model, can provide further insights on model perform-
ance. Furthermore, inexpensive synoptic stream flow
measurements were found to be valuable for model evalu-
ation. Synoptic data allowed an assessment of the spatial
representativity of the model simulation, whereas model
evaluation is usually only focused on the temporal aspects
of runoff dynamics at the catchment outlet. Despite these
apparent model shortcomings of the distributed model con-
cept, it provides a valuable basis for distributed water quality
assessments and will be the subject of further model
development.
ACKNOWLEDGEMENTS
Discharge and precipitation data were provided by SMHI
(Swedish Meteorological and Hydrological Institute). The
authors thank the Department of Aquatic Sciences and
Assessment, Swedish University of Agricultural Sciences,
especially Mats Wallin and Jakob Nisell, for help with the
land-use information. Funding provided by the German
Academic Exchange Service (DAAD) supported an
extended visit of the first author to Sweden.
REFERENCES
Andreassian, V., Perrin, C., Berthet, L., Le Moine, N., Lerat, J.,Loumagne, C., Oudin, L., Mathevet, T., Ramos, M. H. &Valery, A. HESS Opinions ‘Crash tests for astandardized evaluation of hydrological models’. Hydrol.Earth Syst. Sci. 13 (10), 1757–1764.
Arheimer, B. & Brandt, M. Modelling nitrogen transport andretention in the catchments of southern Sweden. Ambio 27(6), 471–480.
Arheimer, B. & Wittgren, H. B. Modeling the effects ofwetlands on regional nitrogen transport. Ambio 23 (6),378–386.
Arheimer, B. & Wittgren, H. B. Modelling nitrogen removalin potential wetlands at the catchment scale. Ecol. Eng. 19(1), 63–80.
Becker, A. & Braun, P. Disaggregation, aggregation andspatial scaling in hydrological modelling. J. Hydrol. 217 (3–4),239–252.
Beldring, S., Engeland, K., Roald, L. A., Saelthun, N. R. & Vokso,A. Estimation of parameters in a distributedprecipitation-runoff model for Norway. Hydrol. Earth Syst.Sci. 7 (3), 304–316.
Bergström, S. The development of a snow routine for theHBV-2 model. Nord. Hydrol. 6, 73–92.
http://dx.doi.org/10.5194/hess-13-1757-2009http://dx.doi.org/10.5194/hess-13-1757-2009http://dx.doi.org/10.1016/S0925-8574(02)00034-4http://dx.doi.org/10.1016/S0925-8574(02)00034-4http://dx.doi.org/10.1016/S0022-1694(98)00291-1http://dx.doi.org/10.1016/S0022-1694(98)00291-1http://dx.doi.org/10.5194/hess-7-304-2003http://dx.doi.org/10.5194/hess-7-304-2003
-
332 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
Bergström, S. Parametervärden för HBV-modellen i Sverige:erfarenheter från modellkalibreringar under perioden 1975–1989. SMHI, Norrköping.
Bergström, S. The HBV model. In: Computer Models ofWatershed Hydrology (V. P. Singh ed.). Water ResourcesPublications, Highlands Ranch, Colorado, pp. 443–476.
Beven, K. J. A discussion of distributed hydrologicalmodelling. In: Distributed Hydrological Modelling (M. B.Abbott & J. C. Refsgaard, eds). Kluwer Academic, Dordrecht,pp. 255–278.
Beven, K. J. Rainfall-Runoff Modelling: The Primer. JohnWiley & Sons, Ltd., Chichester.
Beven, K. A manifesto for the equifinality thesis. J. Hydrol.320 (1–2), 18–36.
Blöschl, G. & Sivapalan, M. Scale issues in hydrologicalmodelling: a review. Hydrol. Proc. 9, 251–290.
Breuer, L., Huisman, J. A., Willems, P., Bormann, H., Bronstert, A.,Croke, B. F. W., Frede, H. G., Graff, T., Hubrechts, L.,Jakeman, A. J., Kite, G., Lanini, J., Leavesley, G., Lettenmaier,D. P., Lindstrom, G., Seibert, J., Sivapalan, M. & Viney, N. R.Assessing the impact of land use change on hydrology byensemble modeling (LUCHEM). I: model intercomparisonwith current land use. Adv. Water Res. 32 (2), 129–146.
Carpenter, T. M. & Georgakakos, K. P. Intercomparison oflumped versus distributed hydrologic model ensemblesimulations on operational forecast scales. J. Hydrol. 329(1–2), 174–185.
Carpenter, S. R., Caraco, N. F., Correll, D. L., Howarth, R. W.,Sharpley, A. N. & Smith, V. H. Nonpoint pollution ofsurface waters with phosphorus and nitrogen. Ecol. Appl. 8(3), 559–568.
Darracq, A., Greffe, F., Hannerz, F., Destouni, G. & Cvetkovic, V. Nutrient transport scenarios in a changing Stockholmand Malaren valley region, Sweden. Water Sci. Technol. 51(3–4), 31–38.
Doherty, J. PEST: Model-independent Parameter Estimation,User Manual. Watermark Numerical Computing, Brisbane.
Doherty, J. & Johnston, J. M. Methodologies for calibrationand predictive analysis of a watershed model. J. Am. WaterRes. Assoc. 39 (2), 251–265.
Donnelly-Makowecki, L. M. & Moore, R. D. Hierarchicaltesting of three rainfall-runoff models in small forestedcatchments. J. Hydrol. 219 (3–4), 136–152.
Eriksson,B. Den ‘potentiella’Evapotranspirationen i Sverige: The‘potential’ Evapotranspiration in Sweden. SMHI, Norrköping.
Exbrayat, J. F., Viney, N. R., Seibert, J., Wrede, S., Frede, H. G. &Breuer, L. Ensemble modelling of nitrogen fluxes: datafusion for a Swedish meso-scale catchment. Hydrol. EarthSyst. Sci. 14, 2383–2397.
Fenicia, F., Savenije, H. H. G., Matgen, P. & Pfister, L. Understanding catchment behavior through stepwise modelconcept improvement. Water Resour. Res. 44 (1), W01402.
Grayson, R. & Blöschl, G. Spatial Patterns in CatchmentHydrology Observations and Modelling. CambridgeUniversity Press, Cambridge, UK.
Grayson, R. B., Moore, I. D. & Mcmahon, T. A. Physicallybased hydrologic modeling, 2. Is the concept realistic. WaterResour. Res. 28 (10), 2659–2666.
Gren, I.-M. The value of investing in wetlands for nitrogenabatement. Eur. Rev. Agric. Econ. 22 (2), 157–172.
Gretener, B. The River Fyris: A Study of Fluvial Transportation.Institute of Earth Sciences, Uppsala University, Uppsala.
Gupta, H. V., Sorooshian, S. & Yapo, P. O. Towardimproved calibration of hydrologic models: multiple andnoncommensurable measures of information. Water Resour.Res. 34 (4), 751–763.
Gupta, H. V., Wagener, T. & Liu, Y. Q. Reconciling theorywith observations: elements of a diagnostic approach tomodel evaluation. Hydrol. Proc. 22 (18), 3802–3813.
Hjulström, F. Studies of the morphological activity of riversas illustrated by the River Fyris. Bulletin of the GeologicalInstitution of the University of Uppsala 25, 221–527.
Hooper, R. P. Applying the scientific method to smallcatchment studies: a review of the Panola Mountainexperience. Hydrol. Proc. 15 (10), 2039–2050.
Hundecha, Y. & Bardossy, A. Modeling of the effect of landuse changes on the runoff generation of a river basin throughparameter regionalization of a watershed model. J. Hydrol.292 (1–4), 281–295.
Hutchinson, M. F. A new procedure for gridding elevationand stream line data with automatic removal of spurious pits.J. Hydrol. 106 (3–4), 211–232.
Jakeman, A. J. & Hornberger, G. M. How much complexityis warranted in a rainfall-runoff model. Water Resour. Res.29 (8), 2637–2649.
Johst, M., Uhlenbrook, S., Tilch, N., Zillgens, B., Didszun, J. &Kirnbauer, R. An attempt of process-oriented rainfall-runoff modeling using multiple-response data in an alpinecatchment, Loehnersbach, Austria. Hydrol. Res. 39 (1), 1–16.
Karssenberg, D., Burrough, P. A., Sluiter, R. & de Jong, K. ThePC raster software and course materials for teachingnumerical modelling in the environmental siences. Trans.GIS 5 (2), 99–110.
Klemes, V. Operational testing of hydrological simulationmodels. Hydrol. Sci. J.–J. Sci. Hydrol. 31 (1), 13–24.
Kotlarski, S. & Jacob, D. Development of a subgrid scaleparameterization of mountain glaciers for use inregional climate modelling. WGNE Blue Book, WMO,Geneva, pp. 4–13.
Krasovskaia, I. & Gottschalk, L. Stability of river flowregimes. Nord. Hydrol. 23 (3), 137–154.
Lantmäteriet. Terrängkartan/Gröna Kartan 1:50.000.Lantmäteriet, Gävle.
Lindgren, G. A., Wrede, S., Seibert, J. & Wallin, M. Nitrogensource apportionment modeling and the effect of land-useclass related runoff contributions. Nord. Hydrol. 38 (4–5),317–331.
Lindström, G., Johansson, B., Persson, M., Gardelin, M. &Bergström, S. Development and test of the distributedHBV-96 hydrological model. J. Hydrol. 201 (1–4), 272–288.
http://dx.doi.org/10.1016/j.jhydrol.2005.07.007http://dx.doi.org/10.1002/hyp.3360090305http://dx.doi.org/10.1002/hyp.3360090305http://dx.doi.org/10.1016/j.advwatres.2008.10.003http://dx.doi.org/10.1016/j.advwatres.2008.10.003http://dx.doi.org/10.1016/j.advwatres.2008.10.003http://dx.doi.org/10.1016/j.jhydrol.2006.02.013http://dx.doi.org/10.1016/j.jhydrol.2006.02.013http://dx.doi.org/10.1016/j.jhydrol.2006.02.013http://dx.doi.org/10.1890/1051-0761(1998)008[0559:NPOSWW]2.0.CO;2http://dx.doi.org/10.1890/1051-0761(1998)008[0559:NPOSWW]2.0.CO;2http://dx.doi.org/10.1111/j.1752-1688.2003.tb04381.xhttp://dx.doi.org/10.1111/j.1752-1688.2003.tb04381.xhttp://dx.doi.org/10.1016/S0022-1694(99)00056-6http://dx.doi.org/10.1016/S0022-1694(99)00056-6http://dx.doi.org/10.1016/S0022-1694(99)00056-6http://dx.doi.org/10.5194/hess-14-2383-2010http://dx.doi.org/10.5194/hess-14-2383-2010http://dx.doi.org/10.1029/2006WR005563http://dx.doi.org/10.1029/2006WR005563http://dx.doi.org/10.1029/92WR01259http://dx.doi.org/10.1029/92WR01259http://dx.doi.org/10.1093/erae/22.2.157http://dx.doi.org/10.1093/erae/22.2.157http://dx.doi.org/10.1029/97WR03495http://dx.doi.org/10.1029/97WR03495http://dx.doi.org/10.1029/97WR03495http://dx.doi.org/10.1002/hyp.6989http://dx.doi.org/10.1002/hyp.6989http://dx.doi.org/10.1002/hyp.6989http://dx.doi.org/10.1002/hyp.255http://dx.doi.org/10.1002/hyp.255http://dx.doi.org/10.1002/hyp.255http://dx.doi.org/10.1016/j.jhydrol.2004.01.002http://dx.doi.org/10.1016/j.jhydrol.2004.01.002http://dx.doi.org/10.1016/j.jhydrol.2004.01.002http://dx.doi.org/10.1016/0022-1694(89)90073-5http://dx.doi.org/10.1016/0022-1694(89)90073-5http://dx.doi.org/10.1029/93WR00877http://dx.doi.org/10.1029/93WR00877http://dx.doi.org/10.2166/nh.2008.035http://dx.doi.org/10.2166/nh.2008.035http://dx.doi.org/10.2166/nh.2008.035http://dx.doi.org/10.1111/1467-9671.00070http://dx.doi.org/10.1111/1467-9671.00070http://dx.doi.org/10.1111/1467-9671.00070http://dx.doi.org/10.2166/nh.2007.015http://dx.doi.org/10.2166/nh.2007.015http://dx.doi.org/10.2166/nh.2007.015http://dx.doi.org/10.1016/S0022-1694(97)00041-3http://dx.doi.org/10.1016/S0022-1694(97)00041-3
-
333 S. Wrede et al. | Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessment Hydrology Research | 44.2 | 2013
Lischeid, G. & Uhlenbrook, S. Checking a process-basedcatchment model by artificial neural networks. Hydrol. Proc.17 (2), 265–277.
Michaud, J. & Sorooshian, S. Comparison of simple versuscomplex distributed runoff models on a midsized semiaridwatershed. Water Resour. Res. 30 (3), 593–605.
Motovilov, Y. G., Gottschalk, L., Engeland, K. & Rodhe, A. Validation of a distributed hydrological model against spatialobservations. Agric. Forest Meteorol. 98 (9), 257–277.
Mroczkowski, M., Raper, G. P. & Kuczera, G. The quest formore powerful validation of conceptual catchment models.Water Resour. Res. 33, 2325–2335.
Nash, J. E. & Sutcliffe, J. V. River flow forecasting throughconceptual models, 1. A discussion of principles. J. Hydrol.10, 282–290.
O’Callaghan, J. F. & Mark, D. M. The extraction of drainagenetworks from digital elevation data. Comput. Vis. Graph.Image Proc. 28, 328–344.
Ott, B. & Uhlenbrook, S. Quantifying the impact of land-usechanges at the event and seasonal time scale using a process-oriented catchmentmodel.Hydrol. Earth Syst. Sci. 8 (1), 62–78.
Perrin, C., Oudin, L., Andreassian, V., Rojas-Serna, C., Michel, C.&Mathevet, T. Impact of limited streamflow data on theefficiency and the parameters of rainfall-runoff models.Hydrol. Sci. J.–J. Sci. Hydrol. 52 (1), 131–151.
Rabus, B., Eineder, M., Roth, A. & Bamler, R. The shuttleradar topography mission – a new class of digital elevationmodels acquired by spaceborne radar. ISPRS J. Photogram.Remote Sens. 57 (4), 241–262.
Refsgaard, J. C. & Henriksen, H. J. Modelling guidelines –terminology and guiding principles. Adv. Water Res. 27 (1),71–82.
Refsgaard, J. C. & Knudsen, J. Operational validation andintercomparison of different types of hydrological models.Water Resour. Res. 32 (7), 2189–2202.
Refsgaard, J. C., Storm, B. & Abbott, M. B. Comment on‘a discussion of distributed hydrological modelling’. In:Distributed Hydrological Modelling (M. B. Abbott & J. C.Refsgaard, eds). Kluwer Academic, Dordrecht, pp. 279–288.
Saelthun, N. R. The ‘Nordic’ HBV model. Description anddocumentation of the model version developed for theproject climate change and energy production. NorwegianWater Resources and Energy Administration Publication 7,Oslo, pp. 26.
Seibert, P. Hydrological characteristics of the NOPEXresearch area. Masters thesis, Department of Earth Sciences(Hydrology), Uppsala University, Uppsala, 51 pp.
Seibert, J. Estimation of parameter uncertainty in the HBVmodel. Nord. Hydrol. 28 (4–5), 247–262.
Seibert, J. Regionalisation of parameters for a conceptualrainfall-runoff model. Agric. Forest Meteorol. 98 (9),279–293.
Seibert, J. & Beven, K. J. Gauging the ungauged basin: howmany discharge measurements are needed? Hydrol. EarthSyst. Sci. 13 (6), 883–892.
Sieber, A. & Uhlenbrook, S. Sensitivity analyses ofa distributed catchment model to verify the model structure.J. Hydrol. 310 (1–4), 216–235.
Seibert, J., Uhlenbrook, S., Leibundgut, C. & Halldin, S. Multiscale calibration and validation of a conceptual rainfall-runoff model. Phys. Chem. Earth B – Hydrol. Oceans Atmos.25 (1), 59–64.
Singh, V. P. Computer Models of Watershed Hydrology. WaterResources Publications, Highlands Ranch, Colorado.
Sooroshian, S. & Gupta, V. K. Model calibration. In:Computer Models of Watershed Hydrology (V. P. Singh, ed.).Water Resources Publications, Highlands Ranch, Colorado,pp. 23–68.
Strasser, U. & Etchevers, P. Simulation of daily discharges forthe upper Durance catchment (French Alps) using subgridparameterization for topography and a forest canopy climatemodel. Hydrol. Proc. 19 (12), 2361–2373.
Uhlenbrook, S.& Sieber, A. On the value of experimental datato reduce the prediction uncertainty of a process-orientedcatchment model. Environ. Model. Softw. 20 (1), 19–32.
Uhlenbrook, S., Roser, S. & Tilch, N. Hydrological processrepresentation at the meso-scale: the potential of adistributed, conceptual catchment model. J. Hydrol. 291(3–4), 278–296.
Uhlenbrook, S., Seibert, J., Leibundgut, C. & Rodhe, A. Prediction uncertainty of conceptual rainfall-runoffmodels caused by problems in identifying model paramtersand structure. Hydrol. Sci. J.–J. Sci. Hydrol. 44 (5),779–797.
Wagener, T., Boyle, D. P., Lees, M. J., Wheater, H. S., Gupta, H. V.& Sorooshian, S. A framework for development andapplication of hydrological models. Hydrol. Earth Sys. Sci.5 (1), 13–26.
Wissmeier, L. & Uhlenbrook, S. Distributed, high-resolution modelling of 18O signals in a meso-scalecatchment. J. Hydrol. 332 (3–4), 497–510.
Xu, C. Y. Operational testing of a water balance model forpredicting climate change impacts. Agric. Forest Meteorol.98 (9), 295–304.
First received 28 June 2010; accepted in revised form 8 September 2011. Available online 21 July 2012
http://dx.doi.org/10.1002/hyp.1123http://dx.doi.org/10.1002/hyp.1123http://dx.doi.org/10.1029/93WR03218http://dx.doi.org/10.1029/93WR03218http://dx.doi.org/10.1029/93WR03218http://dx.doi.org/10.1016/S0168-1923(99)00102-1http://dx.doi.org/10.1016/S0168-1923(99)00102-1http://dx.doi.org/10.1029/97WR01922http://dx.doi.org/10.1029/97WR01922http://dx.doi.org/10.1016/0022-1694(70)90255-6http://dx.doi.org/10.1016/0022-1694(70)90255-6http://dx.doi.org/10.1016/S0734-189X(84)80011-0http://dx.doi.org/10.1016/S0734-189X(84)80011-0http://dx.doi.org/10.5194/hess-8-62-2004http://dx.doi.org/10.5194/hess-8-62-2004http://dx.doi.org/10.5194/hess-8-62-2004http://dx.doi.org/10.1016/S0924-2716(02)00124-7http://dx.doi.org/10.1016/S0924-2716(02)00124-7http://dx.doi.org/10.1016/S0924-2716(02)00124-7http://dx.doi.org/10.1016/j.advwatres.2003.08.006http://dx.doi.org/10.1016/j.advwatres.2003.08.006http://dx.doi.org/10.1029/96WR00896http://dx.doi.org/10.1029/96WR00896http://dx.doi.org/10.1016/S0168-1923(99)00105-7http://dx.doi.org/10.1016/S0168-1923(99)00105-7http://dx.doi.org/10.5194/hess-13-883-2009http://dx.doi.org/10.5194/hess-13-883-2009http://dx.doi.org/10.1016/j.jhydrol.2005.01.004http://dx.doi.org/10.1016/j.jhydrol.2005.01.004http://dx.doi.org/10.1016/S1464-1909(99)00121-5http://dx.doi.org/10.1016/S1464-1909(99)00121-5http://dx.doi.org/10.1002/hyp.5889http://dx.doi.org/10.1002/hyp.5889http://dx.doi.org/10.1002/hyp.5889http://dx.doi.org/10.1002/hyp.5889http://dx.doi.org/10.1016/j.envsoft.2003.12.006http://dx.doi.org/10.1016/j.envsoft.2003.12.006http://dx.doi.org/10.1016/j.envsoft.2003.12.006http://dx.doi.org/10.1016/j.jhydrol.2003.12.038http://dx.doi.org/10.1016/j.jhydrol.2003.12.038http://dx.doi.org/10.1016/j.jhydrol.2003.12.038http://dx.doi.org/10.5194/hess-5-13-2001http://dx.doi.org/10.5194/hess-5-13-2001http://dx.doi.org/10.1016/j.jhydrol.2006.08.003http://dx.doi.org/10.1016/j.jhydrol.2006.08.003http://dx.doi.org/10.1016/j.jhydrol.2006.08.003http://dx.doi.org/10.1016/S0168-1923(99)00106-9http://dx.doi.org/10.1016/S0168-1923(99)00106-9
Distributed conceptual modelling in a Swedish lowland catchment: a multi-criteria model assessmentINTRODUCTIONTHE FYRISÅN CATCHMENTMODEL DEVELOPMENTLumped HBV modelDistributed HBV modelSub-grid variabilitySimulation of channel routing and lakes
DATA BASE AND MODELLING PROCEDURESpatial dataMeteorological and hydrological dataModel parameterization and calibration
RESULTSSplit-sample test resultsProxy-basin test resultsModel comparisonSynoptic runoff measurements
DISCUSSIONSplit-sample test resultsProxy-basin test resultsModel comparisonSynoptic runoff measurements
CONCLUSIONDischarge and precipitation data were provided by SMHI (Swedish Meteorological and Hydrological Institute). The authors thank the Department of Aquatic Sciences and Assessment, Swedish University of Agricultural Sciences, especially Mats Wallin and Jakob Nisell, for help with the land-use information. Funding provided by the German Academic Exchange Service (DAAD) supported an extended visit of the first author to Sweden.REFERENCES